A method for community discovery in heterogeneous networks by integrating deep learning and social recommendation

Abstract

The present invention discloses a heterogeneous network community discovery method integrating deep learning and social recommendation, comprising the following steps: arranging heterogeneous data into a heterogeneous information network through a network model, selecting an effective meta-path, and constructing a user-user social network and a user-community bipartite network; optimizing the user-user social network and the user-community bipartite network through a graph neural network and a labeling mechanism training, and fusing them to construct a user-community extended graph; based on the user-community extended graph, selecting a meta-path, measuring the similarity between a target user and a target community based on the meta-path, and obtaining a target user-target community node pair; inputting the target user-target community node pair into a graph neural network for training, outputting the probability of the target user joining the target community, and applying it to heterogeneous network community discovery. The present invention can perform community recommendation on large-scale social networks and improve the accuracy of community recommendation.

Description

Technical Field

[0001] The present invention relates to the technical field of network recommendation systems, and in particular to a heterogeneous network community discovery method integrating deep learning and social recommendation. Background Art

[0002] In the knowledge society, information sharing and discovery are very important. With the acceptance of the sharing economy and the advancement of information technology, more and more knowledge contribution and discovery activities are shifting from offline to online. Given the advantages of online communities in breaking geographical and time barriers, which facilitates information sharing and discovery, a typical community brings people together around a topic of common interest. In these communities of interest, users can easily interact with each other and share knowledge, ideas or other relevant information. Users have strong motivations to participate in communities due to the need to exchange information, build friendships and establish reputations. Social networking platforms foster communities because they are useful for online advertising, precision marketing and brand management. On the one hand, users in online communities benefit from sharing information with each other, establishing connections with other members and promoting online services. On the other hand, with the rapid development of communities in online social networks, it is difficult for users to choose appropriate communities to join. Therefore, practitioners and researchers have tried to design methods to achieve community discovery.

[0003] Early community recommendation attempts adopted similarity search strategies to recommend communities for users, and then collaborative filtering played an important role in recommendation systems, where users may belong to multiple communities. Most existing techniques require auxiliary information, such as user-item interactions or rating history, and further assume that each user has joined at least one community, which is too idealistic in real scenarios. In addition, due to explicit constraints, the interactions between users and communities are extremely sparse compared to traditional community / item recommendation problems. The recent success of graph neural networks has attracted researchers to design GNN-based methods for recommendation tasks, such as item recommendation and group recommendation, which mainly focus on the user-item bipartite graph and ignore the social interactions between users. It has been proven that social interactions between users are helpful for recommendation systems because users' choices are likely to be influenced by other people around them. Therefore, recent social recommendation methods have further considered the social relationships between users.

[0004] Given its advantages in capturing rich semantics, heterogeneous information networks (HINs) have become a popular tool for analyzing real-world data. HINs represent entities and their relationships as different types of nodes and edges in the network, respectively. It has been introduced into recommendation models such as movies and citations. In order to mine and search information efficiently, meta-paths representing the HIN structure are proposed to capture the semantic relationships between different types of nodes. Meta-paths are simple connections through network patterns and provide guidance for mining HINs. However, HIN mining for recommendation purposes based on meta-paths suffers from information noise or domain knowledge bias. Not every piece of information in HIN is useful, and enumerating all possible meta-paths will bring too much noise to the recommendation. Currently, the selection of meta-paths mainly depends on domain knowledge, so the recommendation performance depends on experts. The selection of effective meta-paths and how to make recommendations effectively and efficiently remain unsolved. Currently, few studies use HINs to model communities and focus on mining the information connections between users and communities to obtain recommendations. Traditional methods find it difficult to model heterogeneous connections between users and communities based on historical data for recommendations. Deep learning provides an opportunity to improve recommendation performance by capturing nonlinear and non-trivial relationships between entities. Summary of the invention

[0005] The purpose of the present invention is to provide a heterogeneous network community discovery method that integrates deep learning and social recommendation with simple algorithm, low resource consumption and high accuracy.

[0006] The technical solution of the present invention is: a heterogeneous network community discovery method integrating deep learning and social recommendation, comprising the following steps:

[0007] Step 1: Arrange heterogeneous data into heterogeneous information networks through network patterns, select effective meta-paths, and construct user-user social networks and user-community bipartite networks;

[0008] Step 2: Optimize the user-user social network and the user-community bipartite network through graph neural network and labeling mechanism training, and fuse the user-user social network and the user-community bipartite network to construct a user-community extended graph;

[0009] Step 3: Based on the user-community extended graph, a meta-path is selected, the similarity between the target user and the target community based on the meta-path is measured, and the target user-target community node pair is obtained;

[0010] Step 4: Input the target user-target community node pair into a fully connected feedforward neural network for training, output the probability of the target user joining the target community, and apply the obtained prediction results to heterogeneous network community discovery.

[0011] Preferably, the expression of the meta-path in step 1 is as follows:

[0012]

[0013] Where: P m is the meta-path, A n+1 is the n+1th type of node, R l It is the lth semantic relationship type.

[0014] Preferably, in step 1, the valid meta-path selected for constructing the user-user social network is specifically a meta-path whose starting node and ending node are both user entities.

[0015] Preferably, in step 1, a user-community bipartite network is constructed, including the following two situations:

[0016] ① If there is a community entity type, then select the meta-path from the starting node as the user entity to the ending node as the community entity to construct a user-community bipartite network;

[0017] ② If there is no community entity type, it is necessary to divide the heterogeneous information network into initial communities and then construct a user-community bipartite network.

[0018] Preferably, if there is no community entity type, it is necessary to divide the heterogeneous information network into initial communities and then construct a user-community bipartite network, including the following steps:

[0019] S1, transforming heterogeneous networks into homogeneous networks;

[0020] S2, select the initial seed node to obtain the initial seed node set;

[0021] S3, based on the initial set of seed nodes, dynamically adjusts the seed nodes in the homogeneous network;

[0022] S4, the adjusted seed nodes and their neighborhoods are used as the initial communities for community division, and then the meta-path from the starting node as the user entity to the ending node as the community entity is selected to construct a user-community bipartite network.

[0023] Preferably, step S3 further includes measuring the consistency of community labels in the neighborhood N(S) of the node set S using the neighborhood label entropy EN(S), and the calculation formula is as follows:

[0024]

[0025]

[0026] Among them, i′ is the category label of the community, l′(u′) is the set of node category labels in S, and δ(·) is the indicator function.

[0027] Preferably, in step 3, the target user is a user to which no community is assigned.

[0028] Preferably, in step 3, the measurement is based on the similarity between the target user and the target community of the meta-path, and the specific steps are as follows:

[0029] Step 3.1: Measure the similarity between the target user and the community users of the target community, specifically:

[0030] ① Determine the meta-path connecting the target user and the community user. The expression is as follows:

[0031]

[0032] ② Use AvgSim to calculate the similarity between the target user and the community users of the target community. The formula is as follows:

[0033]

[0034] in, YesP m The reverse metapath of i is the i-th target user, u j For the jth community user, Label(u i ,u j |P m ) is u i By P m Arrive at u j The probability of Indicates u j pass Arrive at u i The probability of O(u i |R l ) represents u i By R l The set of neighbors reached, |O(u i |R l )| is the size of the neighbor set, O q (u i |R l ) is the qth neighbor in the neighbor set, Sim(u i ,u j |P m ) measures u symmetrically i and u j The similarities between.

[0035] Step 3.2: Based on the similarity between the target user and the community users of the target community, the most similar community users and meta-path P of the target community are used to calculate the target user's meta-path P. m Measuring target user u i and the similarity between the target community c.

[0036]

[0037] Among them, U c,i For the target user u i Aggregation of the most similar community users.

[0038] Preferably, in step 3, AvgSim is used to measure the similarity between the target user and the target community based on the meta-path, specifically:

[0039] ① Determine the meta-path connecting the target user and the target community. The expression is as follows:

[0040]

[0041] ② Calculate the similarity between the target user and the target community. The calculation formula is as follows:

[0042]

[0043] Among them, c j is the jth target community.

[0044] Preferably, in step 4, the prediction results are sorted according to Top-k, a threshold is selected, and the target community with Top-K prediction probability is recommended to the target user for matching, and it is determined whether the target communities overlap, thereby realizing heterogeneous network community discovery.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows: in heterogeneous information networks, by selecting meta-paths, constructing user-user social networks and user-community bipartite networks, the structure of heterogeneous information networks can be accurately expressed, the nonlinear relationship between entities can be captured, and the accuracy of community recommendation can be improved; by training and optimizing user-user social networks and user-community bipartite networks through message passing graph neural networks and labeling mechanisms in deep learning, the deep complex connection relationship between nodes can be captured, and the user-user social network and the user-community bipartite network can be integrated, which can capture the structure of different communities and the potential interactive relationship between users, generate more informative representations, and realize deep interconnection between users; using AvgSim to measure the similarity between target users and target communities based on meta-paths is simpler and more convenient; processing the prediction results through Top-k screening technology is crucial for large-scale data processing and can improve the speed of community recommendation. The present invention can perform community recommendation on large-scale social networks, and improves the accuracy and stability of community recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the operation process of the present invention.

[0047] Figure 2 It is a schematic diagram of the process of constructing a user-user social network and a user-community bipartite network in the present invention.

[0048] Figure 3 This is a schematic diagram of a heterogeneous network model in the present invention.

[0049] Figure 4 The figure is a schematic diagram of the process of constructing a user-community extension graph in the present invention.

[0050] Figure 5 A schematic diagram of the marking mechanism in the present invention for distinguishing different node types.

[0051] Figure 6 It is a schematic diagram of measuring the similarity between the target user and the target community in the present invention.

[0052] Figure 7 It is a schematic diagram of measuring the similarity between a target user and a target community based on a meta-path in the present invention.

[0053] Figure 8 It is a schematic diagram of the process of recommending a community in the present invention. DETAILED DESCRIPTION

[0054] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0055] like Figure 1-8 As shown, the present invention discloses a heterogeneous network community discovery method based on the fusion of deep learning and social recommendation, comprising the following steps:

[0056] Step 1: Figure 2-3 As shown in the figure, the heterogeneous data are arranged into a heterogeneous information network through the network model, and the effective meta-path is selected to construct the user-user social network and the user-community bipartite network. The specific steps are as follows:

[0057] Step 1.1: Given a network model to represent the entity types and relationships involved in heterogeneous data, the data is arranged into a heterogeneous information network according to the network model. Nodes V and edges E represent entity instances and relationship instances in the heterogeneous information network. The entity instances and relationship instances are mapped to the network architecture through the mapping function. The mapping relationship is as follows:

[0058] V→A

[0059] ψ: E→R

[0060] in, is a mapping function that maps entity instances to entity types, including but not limited to users, articles, tags, communities, etc., ψ is a mapping function that maps relationship instances to relationship types, including subordinate relationships between entities, and the number of relationships represents their length.

[0061] Step 1.2: Select valid meta-paths. Meta-paths represent the connection types of entities in heterogeneous information networks. Different meta-paths represent different semantic relationships. The heterogeneous network model has countless meta-paths of different lengths. Among them, long meta-paths contribute little to social recommendations and lead to high computational costs. The present invention limits the length of meta-paths to a certain number, for example, the length of meta-paths is limited to six, so as to select meta-paths from the user to the end of the community, for example, A1 and A2. n+1 The general expression of the meta path is as follows:

[0062]

[0063] Among them, P m is the meta-path, A n+1 is the n+1th type of node, R l It is the lth semantic relationship type.

[0064] Step 1.3: Select a specific meta-path whose starting and ending nodes are both user entities to construct a user-user social network.

[0065] Step 1.4: Construct a user-community bipartite network, including the following two cases:

[0066] ① If there is a community entity type, select the meta-path from the starting node as the user entity to the ending node as the community entity to construct a user-community bipartite network.

[0067] ② If there is no community entity type, it is necessary to divide the heterogeneous information network into initial communities and then construct a user-community bipartite network. The specific steps are as follows:

[0068] S1, transform the heterogeneous network into a homogeneous network, that is, select the user social isomorphic subgraph G whose meta-path starts from the user type node and ends at the user type node.

[0069] S2, divide the local maximum degree point as the initial seed node. The local maximum degree point is the node whose degree is not less than the degree of all nodes in its neighborhood. The set of direct neighbor nodes of a node is called the neighborhood of the node. Considering that small degree nodes in sparse areas cannot become seed nodes, the small degree nodes and their neighborhoods are deleted from the candidate seed set until the candidate seed set is empty and the initial seed node set is selected.

[0070] S3, after obtaining the initial set of seed nodes, dynamically adjust the seed nodes in the homogeneous network. In order to eliminate the impact of the initial seed node selection on the community discovery results, after the seed nodes are expanded to obtain the community, the seed nodes of the community are updated to nodes with high consistency in the neighborhood community belonging within the community. If the nodes in the node neighborhood belong to the same community, the community belonging of the node is relatively certain. Determining nodes with high certainty of community belonging as seed nodes helps improve the quality of community discovery.

[0071] The neighborhood label entropy EN(S) is used to measure the consistency of community labels in the neighborhood N(S) of the node set S. The calculation formula is as follows:

[0072]

[0073] Among them, i′ is the category label of the community, l′(u′) is the set of node category labels in S, and δ(·) is the indicator function.

[0074] Ω={C1,C2,…,C k′} is a community structure of subgraph G, let community be C i′ , assuming that the connection length l′ between nodes does not exceed 3, if · is true, then δ(·) = 1; if · is false, then δ(·) = 0.

[0075] The lower the neighborhood label entropy, the higher the consistency of community labels in the node neighborhood, and the higher the certainty of the node's community belonging. Because in a complex network, large-degree nodes may belong to multiple communities, and community expansion using large-degree nodes as seed nodes may identify multiple communities containing large-degree nodes as one community. Therefore, selecting the node set with the smallest neighborhood label entropy in the community as the adjusted seed set can limit the community range to the effective range, thereby avoiding the generation of overly large communities.

[0076] S4, the adjusted seed node and its neighborhood are used as the initial community for community division. The subgraph G[Seed∪N(Seed)] is run through the label propagation algorithm to obtain the community structure in the local neighborhood Then reselect the meta-path from the starting node as the user entity to the ending node as the community entity to construct the user-community bipartite network.

[0077] Step 2: Optimize the user-user social network and the user-community bipartite network through graph neural network and labeling mechanism training, and fuse the user-user social network and the user-community bipartite network to construct a user-community extended graph.

[0078] like Figure 4As shown in the figure, in the model, the user information and meta-path initialized by the node vector are used as features. Through iterative training of the message passing graph neural network, the information from the neighbors is aggregated to update and optimize the representation of the user node. In the message passing graph neural network, after each user node goes through k layers, the final embedding vector of the user node is obtained. The embedding vector contains all the information from the farther and more distant neighbors. The calculation formula is as follows:

[0079]

[0080] Among them, H (k+1) is the feature vector of the node at the k+1th layer after k layers of convolution, H (k) is the feature vector of the node at the kth layer, W (k) is the parameter of the k-th convolution layer, σ(·) represents the activation function, A′ is the adjacency matrix, I is the identity matrix, Add the A′ matrix after the self-loop to each node, is the value of the element in the pth row and qth column of the matrix, Add 1 to each element in the D matrix. is the exponentiation of the elements on the diagonal, Indicates the number of nodes connected to node x.

[0081] Since users are influenced by adjacent user nodes and community nodes, combining the user-user social network with the user-community bipartite network can learn more information representations from two different perspectives. Each community can be regarded as a subgraph of the user-community extended graph, because the ability to distinguish different subgraphs is important for generating more informative representations of nodes. In this embodiment, the expressive power of the model is enhanced through a labeling mechanism. On the Internet, although the behavior of each user seems random, there are different rules behind these random behaviors. For example: the distribution of user activity and item popularity in the user behavior dataset follows a long-tail distribution. If a defined label is used by a user in a community, the item popularity is 1.

[0082] like Figure 5 As shown, the model distinguishes nodes inside and outside the subgraph by generating different embeddings of non-isomorphic subgraphs by the graph neural network, and generates different node representations according to the information of the nodes in different subgraphs. In this embodiment, the labeling mechanism is specifically to assign a label vector to each node, specify its affiliation information, and define the label of node x given a node x and a community c. The expression is as follows:

[0083]

[0084] A one-dimensional label is reserved for each node. If a node belongs to a community, the label of the node is 1, otherwise the label of the node is 0. A node can be a user node or a community node. Using the labeling mechanism in the message passing graph neural network can enhance the feature representation of the nodes in the model. We can distinguish the neighbor node entity types of nodes in heterogeneous information networks, which is convenient for obtaining different heterogeneous information network subgraphs.

[0085] According to the organizational principle of heterogeneous information networks, user preferences will be affected by the nodes directly connected to them. In this embodiment, the user-user social network and the user-community bipartite network are integrated to construct a user-community extension graph, obtain the final embedding vector of the user node, and propagate the information of the embedding vector in the user-community extension graph, retaining rich user-user interaction information, which is used to generate the embedding vector of the community node.

[0086] Step 3: If Figure 6 As shown, based on the user-community expansion graph, a meta-path is selected, the similarity between the target user and the target community based on the meta-path is measured, and a target user-target community node pair is obtained. The target user is a user to which no community is assigned. The specific steps are as follows:

[0087] Step 3.1: Measure the similarity between the target user and the community users of the target community, specifically:

[0088] ① Determine the meta-path connecting the target user and the community user. The expression is as follows:

[0089]

[0090] ② Use the AvgSim algorithm to calculate the similarity between the target user and the community users of the target community. The formula is as follows:

[0091]

[0092] in, YesP m The reverse metapath of i is the i-th target user, u j For the jth community user, Label(u i ,u j |P m ) is u i By P m Arrive at u j The probability of Indicates u j pass Arrive at u i The probability of O(u i |R l ) represents ui By R l The set of neighbors reached, |O(u i |R l )| is the size of the neighbor set, O q (u i |R l ) is the qth neighbor in the neighbor set, Sim(u i ,u j |P m ) measures u symmetrically i and u j The similarities between.

[0093] Assume that a and b are common variables of target users and community users. If a and b are users of the same type, the value of Label(a,b) is 1, the user is directly connected to itself, and the similarity between the target user and the community user is 1; if a and b are users of different types, the value of Label(a,b) is 0, and the similarity between the target user and the community user is 0.

[0094] Step 3.2: When calculating the similarity between the target user and the target community, all like-minded users of the target user will be considered. If the like-minded users do not belong to the target community and have not joined any community, they will not participate in the calculation of user-community similarity. Based on the similarity between the target user and the community users of the target community, the most similar community users and meta-path P of the target community are used to calculate the similarity between the target user and the community users. m Measuring target user u i and the similarity between the target community c.

[0095]

[0096] Among them, U c,i For the target user u i Aggregation of the most similar community users.

[0097] The greater the similarity between the target user and the target community, the greater the probability that the target user will join the target community. The target user can share all resources in the target community and is more likely to find content that interests him or her.

[0098] If the similarity between the target user and the target community is smaller, the target user may not be able to join the target community and thus cannot share the resources of the target community.

[0099] In another embodiment, by selecting a meta-path, the AvgSim algorithm is used to measure the similarity between the target user and the target community based on the meta-path. The AvgSim algorithm can measure the similarity between entities in a heterogeneous information network through meta-paths of all types and any lengths without adding additional work, and the calculation is more convenient. Specifically,

[0100] ① Determine the meta-path connecting the target user and the target community. The expression is as follows:

[0101]

[0102] ② Calculate the similarity between the target user and the target community. The calculation formula is as follows:

[0103]

[0104] Among them, c j is the jth target community.

[0105] like Figure 7 As shown, taking the target user u2 and the target community c as an example, AvgSim is used to calculate the similarity between the target user u2 and the target community c.

[0106] ① Determine the meta-path connecting the target user u2 and the target community c:

[0107]

[0108] ② Calculate Label(u2,c|UTUC) and Label(c,u2|UTUC -1 ):

[0109]

[0110] ③ Calculate the similarity between the target user u2 and the target community c:

[0111]

[0112] Step 4: Figure 8 As shown, the target user-target community node pair obtained in step 3 is input into a fully connected feedforward neural network DNN with multiple hidden layers for training, and the probability of the target user joining the target community is output. The final prediction result is applied to heterogeneous network community discovery.

[0113] Assume a t is the output vector of the lth layer in the T-layer DNN, a t-1 is the output vector of the t-1th layer in the t-1th layer DNN, a 1 is the input layer, a T is the output layer, and the calculation formula is as follows:

[0114]

[0115] in, is a nonlinear activation function, is the matrix value of the vth column and the fth row of the training weight, is the vth column matrix vector of the bias.

[0116] ReLU is used as the activation function of the hidden layer to avoid gradient disappearance, and Sigmoid function is used as the nonlinear activation function of the output layer to constrain the output value to the range of (0,1), thereby deriving the relationship between the target user and the target community, and recommending the community with Top-K prediction probability to the target user for matching, and determining whether the target community overlaps, thereby realizing heterogeneous network community discovery.

[0117] The present invention constructs a heterogeneous information network according to heterogeneous data, and expresses node information in the form of vectors using a graph neural network, selects effective meta-paths as connectivity information between users, integrates a user-user social network and a user-community bipartite network, and constructs a user-community extension graph; obtains the similarity between the target user and the target community by calculating the similarity between the target user and the community users of the target community in the user-community extension graph, or directly calculates the similarity between the target user and the target community using AvgSim, and obtains a target user-target community node pair; inputs the node pair into a fully connected feedforward neural network for training, obtains a final prediction result, and sorts it according to Top-k, selects a threshold, matches the community with the highest prediction result with the target user of the unassigned community, determines whether the target communities overlap, and realizes heterogeneous network community discovery.

[0118] The present invention makes full use of the advantages of deep learning technology, uses graph neural networks and deep neural networks to solve the community recommendation problem, can perform community recommendations on large-scale networks, and retain the integrity of the data as much as possible. In heterogeneous network community discovery, it combines user-user social networks with user-community bipartite networks, retains the nonlinear connection relationship between users, and can improve the accuracy and stability of community recommendations.

[0119] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A heterogeneous network community discovery method integrating deep learning and social recommendation, characterized by: The steps include: Step 1: Arrange heterogeneous data into heterogeneous information networks through network patterns, select effective meta-paths, and construct user-user social networks and user-community bipartite networks; The construction of the user-community bipartite network includes the following two situations: ① If there is a community entity type, then select the meta-path from the starting node as the user entity to the ending node as the community entity to construct a user-community bipartite network; ② If there is no community entity type, it is necessary to divide the heterogeneous information network into initial communities and then construct a user-community bipartite network, including the following steps: S1, transforms heterogeneous networks into homogeneous networks; S2, select the initial seed node to obtain the initial seed node set; S3, based on the initial set of seed nodes, dynamically adjusts the seed nodes in the homogeneous network; S4, taking the adjusted seed nodes and their neighborhoods as the initial communities for community division, and then selecting the meta-path from the starting node as the user entity to the ending node as the community entity to construct a user-community bipartite network; Step 2: Optimize the user-user social network and the user-community bipartite network through graph neural network and labeling mechanism training, and fuse the user-user social network and the user-community bipartite network to construct a user-community extended graph; Step 3: Based on the user-community extended graph, a meta-path is selected, the similarity between the target user and the target community based on the meta-path is measured, and the target user-target community node pair is obtained; The measurement is based on the similarity between the target user and the target community of the meta-path, and the specific steps are as follows: Step 3.1: Measure the similarity between the target user and the community users of the target community, specifically: ① Determine the meta-path connecting the target user and the community user. The expression is as follows: P m ： ② Use AvgSim to calculate the similarity between the target user and the community users of the target community. The formula is as follows: in, YesP m The reverse metapath of i is the i-th target user, u j For the jth community user, Label(u i ,u j |P m ) is u i By P m Arrive at u j The probability of Indicates u j pass Arrive at u i The probability of O(u i |R l ) represents u i By R l The set of neighbors reached, |O(u i |R l )| is the size of the neighbor set, O q (u i |R l ) is the qth neighbor in the neighbor set, Sim(u i ,u j |P m ) measures u symmetrically i and u j similarities between; Step 3.2: Based on the similarity between the target user and the community users of the target community, the most similar community users and meta-path P of the target community are used to calculate the target user's meta-path P. m Measuring target user u i The similarity between and the target community c is as follows: Among them, U c,i For the target user u i Aggregation of the most similar community users; Step 4: Input the target user-target community node pair into a fully connected feedforward neural network for training, output the probability of the target user joining the target community, and apply the obtained prediction results to heterogeneous network community discovery.

2. According to claim 1, a heterogeneous network community discovery method integrating deep learning and social recommendation is characterized by: The expression of the meta path described in step 1 is as follows: P m ： Where: P m is the meta-path, A n+1 is the n+1th type of node, R l It is the lth semantic relationship type.

3. The method for discovering heterogeneous network communities by integrating deep learning and social recommendation according to claim 1, characterized in that: In step 1, the valid meta-path selected by constructing the user-user social network is specifically a meta-path whose starting node and ending node are both user entities.

4. The method for discovering heterogeneous network communities by integrating deep learning and social recommendation according to claim 1, characterized in that: Step S3 also includes using the neighborhood label entropy EN(S) to measure the consistency of the community labels in the neighborhood N(S) of the node set S. The calculation formula is as follows: Among them, i ′ is the category label of the community, l ′ (u ′ ) is the set of node category labels in S, and δ(·) is the indicator function.

5. The method for discovering heterogeneous network communities by integrating deep learning and social recommendation according to claim 1, characterized in that: In step 3, the target user is a user who is not assigned a community.

6. The method for discovering heterogeneous network communities by integrating deep learning and social recommendation according to claim 1, characterized in that: In step 4, the prediction results are sorted according to Top-k, a threshold is selected, and the target community with Top-K prediction probability is recommended to the target user for matching, and it is determined whether the target communities overlap, thereby realizing heterogeneous network community discovery.

Images